Research On Tensor Deep Computation Model

The arrival of the era of artificial intelligence has spawned a wave of scholars' research on deep learning technology.Deep learning has achieved great success in the fields of target detection,natural language processing,machine translation,and recommendation systems by reasoning analysis and learning summary of existing knowledge and solved difficult problems in many fields.Although the development of deep learning has achieved remarkable merits so far,the traditional deep learning models still have some problems.One is that traditional deep learning models are based on vector space.It is not possible to directly model hybrid heterogeneous data in Cyber-Physical-Social Systems(CPSSs).Instead,these heterogeneous data are converted into vectors for modeling.Although this vectorization method can also work,it does not take the potential correlation between the data into consideration,which results in underutilized the data.The second is that the traditional deep learning models have privacy leakage during the training process.For some sensitive data,once it is leaked,it will bring inestimable threats or property losses to individuals and even the country,even involving life safety.Therefore,this paper conducts research work on the above two issues and the main research contents are as follows:In order to solve the problem that traditional deep learning models cannot directly model multi-source heterogeneous data in CPSSs,this paper uses tensor to represent heterogeneous data and organically combines with traditional deep learning models to propose two kinds of deep computation models.First,based on the traditional recurrent neural network model,this paper combines the tensor with it and proposes a recurrent deep computation model.Second,based on the tensor autoencoder and the deep convolutional computation model,this paper combines the two and adds a tensor feature fusion layer and proposes a hybrid deep computation model.In this paper,a high-order back-propagation algorithm along the time and a hybrid high-order back-propagation algorithm are designed f to train the two proposed models respectively.Aiming at the problem of privacy leakage in the training process of deep computation models,this paper combines differential privacy protection technology on the basis of the proposed recurrent deep computation model,and further proposes a differentially private recurrent deep computation model.In this regard,this paper designs a high-order along-time back-propagation algorithm based on differential privacy and proves that the proposed method satisfies the nature of differential privacy,and also analyzes the usability of the algorithm.Aiming at the three proposed models,experiments were conducted on multiple data sets.The experimental results illustrate that the proposed models and algorithms are effective.